MelangeTry your model ZETIC EXECUTION DESIGN / MOBILE
We build the execution engine for your model.
Then deliver it as an iOS and Android SDK.
MOBILE CUSTOMERS
& ECOSYSTEM
Mobile AI collaboration
iPad SDK delivered & accepted
Model ecosystem partner
Snapdragon Summit showcase01 / THE DIFFERENCE YOUR APP CAN USE
Same phone. A different execution engine.
LFM3-1.7B · tokenized input to first selection
The first token is selected 295.1 ms earlier.
Measured endpointsSame iPhone 17 Pro, same 1,024 input token IDs, 256 token selections. Warm, preloaded Release runs; median of four runs per engine. ZETIC W4A16 versus llama.cpp Metal QAD-Q4_0. Model loading, tokenization and display are outside the first-token timer. Replay illustrates these measured endpoints at 4× slower playback; it is not a live inference run.
Read the research ↗02 / INSIDE MELANGE-BENCH
Real-device evidence. Before your next release.
Choose the model and devices.
Compare the measured paths.
Bring the result into your app.
Generation speed. Best measured run per model across the selected phones.
Real screenshots from the public Melange-Bench. The selected models are LFM2.5-350M, LFM3-1.7B, Hy-MT2-1.8B and Qwen3.5-4B. Speed and first-token bars show the best measured run for each metric under the selected filters, which may be on different phones. Coverage counts selected devices with a run reaching at least 20 tokens/s and at most 2 seconds to first token. These screenshots are a dated product preview, not live results or a claim that every model meets every target. The separate LFM3 research comparison above uses its own reported builds and measurement conditions.
03 / WHAT WE BUILD BELOW THE SDK
We reshape computation, write native kernels,
and redesign how the runtime moves data.
260–267 → 56.8 ms
50-token prompt · Xiaomi 17 Ultra
Native prefill + decode.
The separate QNN context is gone.
We engineer the graph,
the kernels and the runtime
as one execution system.
Specific LFM3 native Hexagon research implementation. Prefill compares a previous fixed 1,024-row QNN graph with native computation of the actual 50 input rows; this is a component timing, not whole-app latency. The 1.77 GB reduction is model-package payload, not RAM. Native prefill and decode share a DSP session and state layout, eliminating the previous per-turn copy/restart. These results do not promise automatic native rewriting or identical gains for every Melange model.
04 / MOBILE WORK. REAL CUSTOMERS.
iPAD / MEDICAL AI SOFTWAREAn on-device SDK integrated into AiDRegen,
with customer acceptance.
SMARTPHONE / CONNECTED APPLIANCESOn-device voice and function calling
for ThinQ-connected appliances.
05 / YOUR MODEL TO YOUR NEXT RELEASE
Upload supported ONNX or PyTorch Exported Program (.pt2) models with fixed-shape sample inputs, or start from a supported library model. Melange supplies hardware-specific builds and a device-aware runtime. Apple and Qualcomm paths include GPU/CPU fallback; Tensor, Dimensity and Exynos NPU support is enterprise scoped. Downloads are cached on the device; inference runs locally. Custom native engine work is scoped to the model and target.
Check compatibility ↗ · SDK quickstart ↗YOUR MODEL. YOUR PHONES. YOUR PERFORMANCE GOALS.